DADOS, DA CORRELAÇO ESPAÇO-TEMPORAL
E CONSUMO DE ENERGIA PARA REALIZAR
SOLUÇÕES CIENTES DE AGREGAÇO DE
DADOS, DA CORRELAÇO ESPAÇO-TEMPORAL
E CONSUMO DE ENERGIA PARA REALIZAR
COLETA DE DADOS EM REDES DE SENSORES
SEM FIO
Tese apresentada ao Programa de
Pós--GraduaçãoemCiênia daComputaçãodo
Instituto de Ciênias Exatas da
Universi-dade Federal de Minas Gerais omo
requi-sito parial para a obtenção do grau de
Doutor emCiênia daComputação.
Orientador: Antonio Alfredo Ferreira Loureiro
Co-orientadora: Regina Borges de Araujo
DATA AGGREGATION, SPATIO-TEMPORAL
CORRELATION AND ENERGY-AWARE
SOLUTIONS TO PERFORM DATA COLLECTION
IN WIRELESS SENSOR NETWORKS
Thesis presented to the Graduate Program
inComputerSiene ofthe Federal
Univer-sityofMinas Geraisinpartialfulllmentof
the requirements for the degree of Dotor
inComputer Siene.
Advisor: Antonio Alfredo Ferreira Loureiro
Co-advisor: Regina Borges de Araujo
Primeiramente aos meus pais, Antnio Villas Martins e TerezinhaJoana Gonçalves
Villas, peloamor, inentivo e dediação, sempre areditando no meu suesso.
Certamente orgulhosos por mais umimportante passo na minha vida;
À minha irmã,Daiane Villas, que sempre toreu para que tudo desse erto;
À Vernia, minha namorada e possivelmente futura esposa, pelapaiênia nos
momentos mais difíeis;
Ao meu querido sobrinho Geovane Villas, ao qual tenho um amor imenso;
Ao ApareidoGonzales Castilho, meu primo, mesmo não estando presente entre nós,
foi um exemplo de vida para mim e me ensinou que nos estudos eu onseguiria
ADeus, aimadetudoetodos,pelaoportunidadedeaperfeiçoamentotantointeletual
quanto moral. Pelapresença onstante emada momento de minhavida, pelafamília
e pelos amigos;
A minha família (Antonio, Terezinha, Daiane e Geovane), pessoas sem as quais
eu não seriaoquesou hoje. Mesmo distantes siamente, permaneerampróximosem
pensamento atravésde suas orações, onversas e onselhos;
Ao meu orientador, Prof. Antonio Alfredo Ferreira Loureiro, pela ajuda,
inen-tivo,eslareimentoe,aimadetudo,exemplo. Aminhao-orientadoraReginaBorges
de Araújo pela motivação, enorajamento, amizade e direionamento. Ao meu
super-visor,Prof. AzzedineBoukerhe pelaoportunidadeedireionamentoduranteoestágio
de sanduíhe realizado na University of Ottawa. Ao Prof. Horáio A. B. Fernandes
de Oliveira pelas disussões e olaborações e ao meu amigo Daniel L. Guidoni pelas
disussõese olaboraçõesem vários trabalhos;
AosamigosdoDCC-UFMG,Alyson,Celso,CésarSoares,DanielGalinkin,Daniel
Guidoni, Eduardo Muelli, Felipe, Fernanda, Fernando, Flávio, Guilherme, Heitor,
Izabela, John Holiver, Laura, Letíia, Marelo, Max, Pedro, Rafael Colares, Rafael
Santin,ThiagoPedpano,TiagoCunhaeZiltonpelaamizade,ompanhiaeajudadiária;
Ao pessoal do laboratório Ubiquitous and Wireless Sensor Networks Lab pelo
ambientedesontraído e,ao mesmotempo,prossional que enontrei no laboratório;
Aos meus amigos do Paradise Laboratory, Aissa, Cristiano, Daniel, Heitor,
Riharde Robsonpelaompanhia,amizadee apoiodesde minhahegada emOttawa;
A todos os meus familiarespelaonança irrestrita;
Aosmeusamigosde infânia,que sempreestiveramaomeu lado,mesmoquando
Este trabalhoofereeuma disussãogeralsobre otemade agregaçãode dadose
explo-ração da orrelação espaço-temporal dos dados em redes de sensores sem o(RSSFs)
que permite: (i) a identiação de problemas em aberto e (ii) o entendimento dos
re-quisitos eimpliaçõesdouso de agregação de dadosemRSSFs, alémdaexploraçãoda
orrelação espaço-temporal dos dados.
Esta disussãoéfeitaatravésde umlevantamentobibliográodoestado-da-arte
envolvendo agregação e orrelação espaço-temporalde dados em RSSFs. Como
resul-tado daanálise de arquiteturas, modelos emétodos de agregação e orrelação
espaço-temporal de dados identiados neste levantamento bibliográo, propomos quatro
soluções diferentes para o problema de agregação e exploração da orrelação
espaço-temporalde dadosonsiderandodiferentes enáriosemRSSFs: osalgoritmosDAARP,
DDAARP, DST e EAST. Os algoritmos propostos reduzem o número de mensagens
neessárias para riar uma árvore de roteamento, maximizam o número de rotas
so-brepostas, seleionam as rotas om maior taxa de agregação e realizam transmissões
onáveis de dadosagregados.
Assoluçõespropostasforamamplamenteomparadasomoutrassoluçõesda
lite-raturaemrelaçãoaos ustosdeomuniação,eiêniadeentrega,taxade agregaçãoe
taxade entrega de dadosagregados. Osresultadosmostramqueassoluçõespropostas
podem ser uma boa alternativa para agregar dados e explorar a orrelação
espaço-temporaldos dados durante oroteamento. Diversos experimentossão mostrados para
OdoumentodestateseestáredigidoeminglêsomotítuloDataAggregation,
Spatio-TemporalCorrelationandEnergy-AwareSolutionstoperformDataColletion in
Wire-less Sensor Networks. Para atender às normas da Universidade Federal de Minas
Gerais, este resumo em português faz um resumo estendido de ada apítulo desta
tese.
Capítulo 1 Introdução
Os reentes desenvolvimentos nas áreas de omuniação sem o e sensores
multifun-ionaisom apaidadede omuniaçãoeproessamentoimpulsionaramoresimento
dasredesdesensoressemo(RSSFs). AsRSSFsestãoadavezmaispresentesem
apli-açõesomomonitoramentoambiental,vigilâniade amposmilitaresemuitas outras
ondeapresença humananãoépossívelounãodesejada. Umnósensor,porsisó,
apre-senta uma apaidade limitada de deteção de sensoriamento de uma dada grandeza,
mas aapaidadeglobalde deteçãopodeser aumentadadrastiamentequandoosnós
são ombinados formando uma rede de sensores sem o. Logo, nós sensores em uma
RSSF podem monitorar ooperativamente uma determinada área de interesse. Por
exemplo, se oorrer um vazamento de gás em uma sala repleta de botijões de gás e
existir apenas um nó sensor nessa sala, será possível apenas dizer se há ou não
vaza-mento de gás. Por outro lado, se for utilizada uma RSSF adequadamente projetada,
serápossívelnãosódetetarovazamento,masindiarondeovazamentoiniioueomo
ele evoluiu. Um monitoramento dessa forma pode salvarvidas e patrimnio, além de
diminuirustos de seguros.
Osnóssensoressãodispositivostipiamenteomrestriçõesdeenergia. Oonsumo
de energiaé geralmente assoiado àquantidade de dadostransmitidos na rede, pois a
omuniação é a atividade que tende a demandar uma maior quantidade de energia.
Uma soluçãosimplespara esse problemaseria areposição dabateriados nós sensores.
vulões ou do espaço. Dessa forma, algoritmos e protoolos projetados para uma
RSSF devem onsiderar o onsumo de energia emsua onepção. Além disso, os nós
sensores podemoletaruma grandequantidade dedadosque preisamser proessados
eenaminhadas,muitas vezes usandoomuniação multihop,emdireçãoaumnósink,
oqualfunionaomoumgateway paraumaestaçãode monitoramento. Nesse enário,
oroteamentodesempenha um papelmuitoimportantenoproesso deoleta de dados.
Pararealizaraoletadedadosdeformamaiseienteeeazomumusomínimo
dereursoslimitados,nóssensoresdevemseronguradosparareportardadosdeforma
inteligente tomando deisões loais. Para isso, a agregação de dados e a exploração
da orrelação espaço-temporal de dados são ténias eazes de eonomia de energia
em RSSFs. Devido à redundânia dos dados brutos reolhidos pelos nós sensores, a
agregação de dados e a orrelação espaço-temporalde dados muitas vezes podem ser
usadas para diminuir o usto de omuniação, eliminando a redundânia de dados e
reportando apenas informações agregadas.
A agregação de dados tem sido utilizada em RSSFs om dois propósitos: (i)
tirar proveito da redundânia emelhorara preisão dos dados; e (ii)reduzir otráfego
de dados e eonomizar energia. No entanto, as propostas atuais têm um usto alto
para riar estruturas de roteamento ientes de agregação de dados e muitas delas não
onsideramaorrelaçãoespaço-temporaldedados. Alémdisso,amaioriadaspropostas
não lidaom falhas nos nós e interrupções nas omuniações, o que provoa perda de
dadose não garantea entrega dos dados oletados.
A prinipal ontribuição desta tese é o desenvolvimento de quatro diferentes
soluções ientes de agregação de dados, da orrelação espaço-temporal e onsumo de
energia para oleta de dados emRSSFs, que nos referimos omo DAARP, DDAARP,
DSTe EAST, quaisserão apresentados, respetivamente, nos apítulos3, 4,5 e 6.
Capítulo 2 Fundamentação Teória
Redes de Sensores sem Fio
UmaRededeSensoressem o(RSSF)podeserdenida omoumaredeooperativade
nóssensores semo, operadostipiamenteporbateria,ujoprinipalobjetivoéduplo:
monitoraro ambiente e transmitir os dados reolhidos para um nó sorvedouro (sink)
usando normalmente omuniação multihop. Este nó sorvedouro será responsável por
proessar todos os dados reebidos dos nós fontes e reportá-los para uma estação de
analisado. Geralmente, ada nó sensor é equipado om vários tiposde sensores omo,
porexemplo, temperatura, pressão, sísmio, aústio, radiação einfravermelho. Esses
nós sensores são onstruídos para serem baratos e normalmente possuem limitações
omputaionais, memória,omuniação e energia.
As RSSFs possibilitamaoleta de informaçõesneessárias emambientes onde o
uso de os ou abeamento não seja possível ou viável. Elaspodem estar inseridas na
estrutura de um prédio, ponte, no interior de máquinas, tubulações, dentro de asas,
orestas, áreas de desastre, plantações, vulões, ampo debatalhaeaté mesmodentro
do orpo humano omo, por exemplo, a retina. O baixo usto dos nós sensores e o
potenialdessa tenologia justiam asua utilizaçãoem diversas áreas, taisomo:
Saúde: ontrole de doenças ontagiosas; interfae para deientes;
monitora-mento de paientes; diagnóstio de distúrbios; administraçãode drogas em
hos-pitais, monitoramento eloalizaçãode paientes emédiosem hospitais.
Apliações Militares: monitoramento de tropas; reonheimento de terreno;
deteção de alvos e de ataques biológios,químios ounuleares.
Meio-ambiente: rastreamento do movimento dos pássaros, pequenos animais;
monitoramentodeondiçõesambientaisqueafetamolheitaseplantio(por
exem-plo,ombateàgeada,deteçãodeomponentesquímiosoubiológios,irrigação);
mapeamentodabio-omplexidadeambiental,estudodapoluiçãoemuitasoutras.
Monitoramento de estrutura/equipamentos: monitoramento e
identi-ação de falhas em estruturas (pontes, prédios, et); monitoramento da fadiga
de máquinas e equipamentos (motores, dutos de gás, et); diagnóstios de
má-quinas.
Apliações omeriais: automação de vendas e proesso industriais;
manutenção de inventário, monitoramento de qualidade de produtos; deteção
e vigilânia de veíulose estabeleimentos.
A tendênia é a produção dos nós sensores em larga esala, barateando o seu
ustoeoinvestimentonodesenvolvimentotenológiolevandoanovasmelhorias,omo
aumentode proessamento e armazenamentoe redução do tamanhodos nós sensores.
Portanto,novasapliaçõespodemsurgiraumentandoaabrangêniadeuso dasRSSFs.
A posição de ada nósensor não preisa ser neessariamente pré-determinada, o
redesde sensores devempossuir a araterístiade auto-organizaçãodos nós.
Agregação de Dados no Roteamento
Agregação de dados durante o roteamento em redes de sensores sem o envolve
dife-rentes formasde transmitirpaotesde dados am de ombinardados provenientes de
fontes diferentes, mas destinado a um nó espeío hamado sink. Um
omponente-have para a agregação de dados em RSSFs é um protoolo de roteamento iente de
agregaçãode dadosbemonebido,quedeterminaondeaagregaçãodeveserrealizada.
Agregaçãodedadosrequerumparadigmadeenaminhamentodiferentedoroteamento
lássio. Protoolosde roteamentolássiotipiamenteusam osaminhos mais urtos
emrelaçãoaalgumamétriaespeía paratransmitirdadosaosink. Emprotoolos
de roteamento iente de agregação de dados, os nós devem enaminhar os paotes de
dados om base no onteúdo do paote e esolher o próximo hop que maximiza a
so-breposiçãoderotasparapromoveraagregaçãodedadosnarededuranteoroteamento.
Para realizar a agregação de dados na rede é neessária alguma forma de
sin-ronizaçãoentre osnós. Tipiamenteosnós não enviam dados,logoque sejapossível.
A espera de informações provenientes de nós vizinhos pode levar a melhores
oportu-nidadesdeagregaçãodedadose,onsequentemente, melhordesempenho. Asprinipais
estratégiasde temporização propostas na literaturasão resumidas a seguir:
Periodi simple aggregation exige que ada nó espere por um períodode tempo
pré-denido,paraagregartodosospaotesdedadosreebidosduranteessetempo
pré-denido e, em seguida, envia um paote de dados om o resultado da
agre-gação de todos os paotesde dados reebidos;
Periodi per-hop aggregation é bastante semelhante à abordagem anterior. A
únia diferença é que o paote om os dados agregados é transmitido logo que
o nó reebe um paote de dados de todos os seus lhos. Isto requer que ada
nó onheça os seus lhos. Além disso, um tempo limite é usado em aso de um
paote de dados de algumlho ser perdido durantea transmissão.
Periodi per-hop adjusted aggregation ajusta o tempo de espera de um nó,
de-pendendo daposiçãodo nóna estrutura de roteamento.
É importante observar que a esolha da estratégia de tempo afeta fortemente a
taxa de agregação em rede e a latênia para relatar os dados oletados. Se o tempo
menor.
Correlação Espaço-Temporal de Dados
Podemos enontrar atualmente na literatura três ategorias prinipais de protoolos
ientes de orrelação de dados: (i) orrelação espaial; (ii) orrelação temporal e (iii)
orrelação espaço-temporal. A seguir, apresentamos os benefíios da exploração da
orrelação espaial/temporalde dados emRSSFs:
1. Correlação Espaial: nós espaialmente próximos tendem a sensoriar valores
semelhantes. No entanto, essa proximidade depende dos requisitos da apliação
e araterístias do evento. Algumas apliações são mais rítias e são menos
tolerantes a disrepânias nos valores sensoriados sobre o fenmeno observado,
exigindoque nóspróximosreportamosdadossensoriados (regiãodeorrelação é
menor). Poroutrolado,outrasapliaçõespodemsermaistolerantesa
disrepân-ias nos valores sensoriados, não exigindo que nós próximos reportam os dados
sensoriados (região de orrelação é maior).
Região de orrelação: emumaregiãode orrelação,osvalores sensoriados
pe-los nós sensores são onsiderados semelhantes para a apliação e, portanto, uma
únia leitura dentro dessaregiãoé osuiente para representá-la. Otamanho da
região de orrelação varia de apliação para apliação e de evento para evento.
Assim, o tamanho da região de orrelação está diretamente relaionado à
apli-ação.
2. Correlação Temporal: tipiamente, a leitura feita pelos sensores no ambiente é
periódia. Consequentemente, os dados sensoriados onstituem uma série
tem-poral. Devido à natureza do fenmenofísio, háuma orrelação temporal
signi-ativaentre adaobservaçãoonseutivade um nósensore osdadosreolhidos
são geralmentesemelhantes emumurtoperíodode tempo. Assim, nessesasos,
os nós sensores não preisam transmitir suas leituras se a leitura atual estiver
dentrode um limiaraeitávelem relaçãoà última leiturareportada.
Correlação temporal: adanófonte mantéma últimaleiturareportada(
R
old
).Quando a leitura atual (
R
new
) estiver disponível,R
new
é omparada omR
old
.R
new
de um nó fonte é reportada se um dado limiar é maior que a tolerâniana oerênia temporal (tt), isto é,
|
(
R
new
−
R
old
)
|
R
old
×
100
> tct
, ondetct
é apor-entagem de tolerânia na oerênia temporal. Caso ontrário o valor
R
new
éapresenta tanto a orrelação espaial quanto a temporal, ou seja, nós
espaial-mentepróximostendemasensoriarvaloressemelhanteseosdadosreolhidossão
geralmente semelhantes em um urto períodode tempo. Neste aso, assoluções
que utilizam ambas as orrelações podem tirar proveito da natureza do evento
detetado e reduziro número de dados reportados.
Na literatura existente de algoritmos que exploram a orrelação espaial e/ou
temporaldos dados,amaioriadasabordagens propostas nãoonsideraonívelde
ener-giados nós na seleçãodos nós representativose asaraterístias doevento durante a
oletadedadosparamelhoresolherosnósrepresentantesdeadaregiãodeorrelação.
As abordagens que onsideram o nível de energia dos nós apresentam um alto usto
deontrole egeralmenteresultamemaltos atrasosedados desatualizadossão
enami-nhados para o nó sink. No entanto, elas não exploram a orrelação espaço-temporal
eientemente.
Capítulo 3 DAARP: Um Protoolo de Roteamento
Ciente de Agregação de Dados para RSSFs
É um novo protoolo de roteamento iente de agregação de dados para RSSFs. A
prinipalmotivaçãoparaprojetarumnovoalgoritmoderoteamentoientedeagregação
dedadoséqueassoluçõesdaliteraturaapresentamum altoustoparariarestruturas
deroteamentoientesdeagregaçãodedados. OalgoritmoDAARPonstróiumaárvore
de roteamento om os aminhos mais urtos (em saltos) que oneta todos os nós
fontes ao sink enquanto maximizaa agregaçãode dados,uja prinipalontribuiçãoé
maximizaraagregação de dadosaolongodarota de omuniação,de uma formamais
onável,atravésdeummeanismode enaminhamentotoleranteafalhas. Resultados
de simulações(apresentados naseção3.6)revelam queo DAARP tem algunsaspetos
haves exigidos pela agregação de dados em RSSFs omo um número reduzido de
mensagensparaariaçãodeumaestruturaderoteamento,maximizaonúmeroderotas
Roteamento Dinâmio e Ciente de Agregação de
Dados para RSSFs
ÉumnovoprotooloderoteamentoientedeagregaçãodedadosdinâmioparaRSSFs,
que usa o nósink para oproessamentoe onguraçãodas rotas ientes de agregação
de dados. A prinipalmotivação para projetaruma abordagemdinâmia,que ria
es-truturas deroteamentodinâmiasientes deagregaçãode dados,équeaqualidadedas
estruturas de roteamentoriadaspeloDAARP etambémpelamaioriados algoritmos
naliteraturadepende daordemde oorrêniados eventos. Assim, umavez riadaessa
estrutura, as rotas são mantidasestátias durantea oorrênia doevento. A prinipal
ontribuiçãodoDDAARPéqueasrotasriadasnãodependemdaordemdeoorrênia
dos eventos e não são mantidas xas durante a oorrênia de eventos. Resultados de
simulações(apresentadosnaseção4.6)revelam queoDDAARP apresenta baixousto
em termos de paotes de ontrole, melhoraa qualidade daestrutura de roteamento e
maximiza a agregação de dados ao longo da rota de omuniação de uma formamais
onável, através de um meanismo de roteamentotolerante afalhas.
Capítulo 5 DST: Um Protoolo de Roteamento
Esalável, Dinâmio e Ciente de Agregação de
Dados para RSSFs
É um novo protoolo de roteamento iente de agregação de dados que leva a uma
so-lução esalável,dinâmiae apresentabaixousto para riarestruturas de roteamento.
Apesar do DDAARP ter mostrado bons resultados, elesofre om problemas de
esal-abilidade e torna-seinviável para redes de larga esala. Alémdisso, onó sink preisa
de onheimento global da rede. DST é uma solução eiente de agregação de dados
que permite o roteamento esalável e dinâmio em RSSFs, que onstrói estruturas de
roteamento om as rotas mais urtas (distânia Eulidiana) que oneta todos os nós
fontes ao nó sink maximizando a agregação de dados e reduzindo a distânia para
onetar ada nó fonte aosink. Resultados de simulações(apresentados naseção 5.6)
revelamqueoDSTapresentabaixoustoemtermosde paotesde ontrole, maximiza
os pontosde agregação e melhoraa qualidade da estrutura de roteamento ofereendo
Eiente para Coleta de Dados Ciente da Correlação
Espaço-Temporal para RSSFs
ÉumnovoalgoritmoientedeenergiaparaoenaminhamentodedadosemRSSFsque
aproveita os meanismos de orrelação espaiale temporalpara eonomizar energia e
manter o enaminhamento de dados preisos em tempo oportuno para o nó sink. A
prinipalmotivaçãopara aonepção doEAST équeamaioriados algoritmosientes
de orrelação espaial e/ou temporal não onsidera o onsumo de energia durante a
oletade dadosparamelhoresolherosnósrepresentativos. Alémdisso,essassoluções
possuemumgrande númerode mensagensde ontroleenão explorade formaeiente
a orrelação espaço-temporal dos dados, nem a sua dinamiidade. A prinipal
on-tribuiçãoéum meanismo de orrelação espaço-temporalde dados emque osnós que
detetaramo mesmoeventosão dinamiamenteagrupados emregiõesorrelaionadas
eum nórepresentanteéseleionadoemadaregiãodeorrelaçãoparaobservaro
fen-meno. Resultados de simulações(apresentadosnaseção 6.4)mostramlaramenteque,
usandoambas asorrelaçõesespaialetemporal,asinformaçõessobreoeventopodem
ser sentidas om altapreisão eainda eonomizar aenergia residualdos nós.
Capítulo 7 Conlusões
Estatese estudou aimportâniaderealizaragregaçãode dadoseexplorara orrelação
espaço-temporaldos dados duranteo roteamento. Devido à impossibilidade de se ter
umaúniasoluçãoparaumdeterminadoproblemaemRSSFs,nestatesesão propostos
quatro algoritmos diferentes para a agregação de dados e exploração da orrelação
This work provides a general disussion for data aggregation that exploits
spatio-temporal data orrelation in wireless sensor networks (WSNs), allowing us to
iden-tify open issues and understand the requirementsand the impliationsregarding data
aggregation, and spatio-temporaldata orrelationin WSNs.
In this disussion, we survey the state-of-the-art of data aggregation and
spatio-temporaldataorrelation inWSNs. Byassessingthe arhitetures, models,and
meth-ods of data aggregation and spatio-temporaldata orrelation identied in the survey,
we propose four dierent solutions for the data aggregation and spatio-temporaldata
orrelation that are suitable for dierent senarios in WSNs. The proposed solutions
are alled DAARP, DDAARP, DST and EAST. The proposed algorithms redue the
number of message neessary to set up a routing tree, maximize the number of
over-lapping routes, selet routes with the highest aggregation rate, and perform reliable
data aggregation transmission.
The proposed solutions have been extensively ompared with other solutions in
the literature and the results show that the proposed solutions may be potential
al-ternatives to perform data aggregation and spatio-temporal data orrelation during
the routing proess. We also present an extensive set of experiments to evaluate the
performane of our algorithms. Our results indiate that our proposed solutions are
suitable for implementationin WSNs.
Keywords: WSNs, routing algorithms, data aggregation, spatio-temporal
2.1 Data routinginWSNs [Oliveira et al.,2009℄. . . 8
2.2 Data Aggregation Aware Routing . . . 9
2.3 Energy onsumption of nodes during the data olletion when using (a) a
lassialapproah; and when using (b) a spatialorrelation based approah. 17
2.4 (a)The time series and (b) The pieewise linear presentation . . . 19
3.1 Example of establishingnew routesand updatingthe hop tree . . . 27
3.2 Example of path repair . . . 29
3.3 Number of Steiner nodes in the routing tree built by the DAARP, InFRA,
MST, and SPT algorithms . . . 34
3.4 Impat of the Network Size . . . 36
3.5 Impat of the Number of Events . . . 37
3.6 Impat of the Event Duration . . . 38
3.7 Impat of Communiation Failures . . . 39
4.1 Overhead . . . 50
4.2 Eieny . . . 52
4.3 Number of Steinernodes . . . 53
5.1 Examples of routingstruture establishmentfor DST variations . . . 59
5.2 Impat of Event Sale . . . 62
5.3 Impat of Network Sale . . . 63
5.4 Impat of Event Duration . . . 64
6.1 Examples of routingstruture used by the EAST algorithm. . . 69
6.2 SpatialCorrelation Mehanism appliedto the event area . . . 72
6.3 Data olleted fromAmazon rainforest. . . 75
6.4 Number of representative nodes . . . 76
6.7 Auray in the readingsof min value. . . 80
6.8 Auray in the readingsof max value . . . 80
6.9 Auray in the readingsof mean value . . . 80
6.10 Notiations
×
Readings . . . 836.11 Average energy onsumption . . . 84
6.12 Data auray. . . 85
6.13 Average delayin high reporting rate. . . 86
2.1 Summaryofthebasiharateristisofsomedataaggregationawarerouting
protools . . . 12
2.2 Summary of the basi harateristis of the main proposed spatial and/or
temporaldata orrelations algorithmsfor WSNs . . . 16
3.1 Communiation omplexity of assessed algorithms . . . 30
3.2 Summaryof the basiharateristisof assessed algorithms. . . 31
3.3 Simulationparameters . . . 32
3.4 Senario with 6 events, 256 nodes and density 20 . . . 35
3.5 Senario with 6 events, 256 nodes and density 30 . . . 35
3.6 Senario with 6 events, 2048 nodes and density 20 . . . 35
3.7 Senario with 6 events, 2048 nodes and density 30 . . . 35
4.1 Communiation omplexity of assessed algorithms . . . 48
4.2 Simulationparameters . . . 49
5.1 Communiation omplexity of assessed algorithms . . . 60
5.2 Simulationparameters . . . 61
6.1 Simulationparameters . . . 74
1 Buildingthe hop tree . . . 25
2 Cluster formationand leader eletion . . . 26
3 Route formation, hop tree updates and data transmission . . . 27
4 Buildingthe hop tree . . . 43
5 Colleting informationabout nodes' position. . . 44
6 Cluster formationand leader eletion . . . 45
7 Routing formation . . . 46
8 Data Transmissions . . . 47
9 Disovery of neighbors' position . . . 56
10 Cluster formationand leader eletion . . . 57
Agradeimentos xi
Resumo xiii
Resumo Estendido xv
Abstrat xxiii
List of Figures xxv
List of Tables xxvii
1 Introdution 1
1.1 Motivation . . . 1
1.2 Objetives . . . 2
1.3 Main Contributions . . . 3
1.4 Organization ofthe Thesis . . . 5
2 Bakground 7
2.1 Wireless Sensor Networks . . . 7
2.2 In-network Data Aggregation . . . 9
2.2.1 Routing Sheme Aware Data Aggregation . . . 11
2.3 Exploiting Spatio-TemporalCorrelation . . . 15
2.3.1 Spatial Correlation . . . 15
2.3.2 TemporalCorrelation . . . 18
2.3.3 Spatio-TemporalCorrelation . . . 20
2.4 Final Remarks. . . 21
3 DAARP: Data Aggregation Aware Routing Protool 23
3.3 Routing Formation,HopTree Updates and Data Transmission . . . 25
3.4 Route Repair Mehanism . . . 28
3.5 Complexity Analysis . . . 29
3.6 Performane Evaluation . . . 31
3.6.1 Methodology . . . 31
3.7 Final Remarks onDAARP . . . 39
4 DDAARP: Dynami Data Aggregation Aware Routing Protool 41
4.1 Buildingthe Hop Tree and Gathering InformationAbout Nodes' Position 42
4.2 Cluster Formationand Leader Eletion . . . 44
4.3 Routing Formation . . . 45
4.4 Data Transmission . . . 46
4.5 Complexity Analysis . . . 47
4.6 Performane Evaluation . . . 48
4.6.1 Simulation Senario and Metris Used . . . 48
4.6.2 Simulation Results . . . 49
4.7 Final Remarks onDDAARP . . . 54
5 DST: Dynami and Salable Tree 55
5.1 Disovery of Neighbors' and Sink's Positions . . . 56
5.2 Cluster Formationand Leader Eletion . . . 57
5.3 Notiationof a New Event . . . 57
5.4 Routing Tree Creation and DataTransmissions . . . 58
5.5 Complexity Analysis . . . 60
5.6 Performane Evaluation . . . 60
5.6.1 Methodology . . . 60
5.7 Final Remarks onDST . . . 65
6 EAST: Eient Data Colletion Aware of Spatio-Temporal
Corre-lation 67
6.1 Spatial CorrelationModel . . . 67
6.2 TemporalCorrelation Model . . . 68
6.3 Overview of the EAST Algorithm . . . 68
6.3.1 Node Loalization. . . 70
6.3.2 Cluster Formation, Leader Eletion, and Division of the Event
AreaintoCells . . . 70
6.4.1 Methodology . . . 74
6.4.2 Event Model. . . 74
6.4.3 Performane Evaluationof the SpatialCorrelation Mehanism . 75
6.4.4 Performane Evaluationof TemporalCorrelation Mehanism . . 81
6.5 Final Remarks onEAST . . . 87
7 Final Remarks 89
7.1 Conlusions . . . 89
7.2 Limitations . . . 91
7.3 Diretions for FutureResearh . . . 91
7.4 Comments onPubliations . . . 92
7.4.1 Journals . . . 92
7.4.2 Conferenes . . . 93
Introdution
1.1 Motivation
Reent developments in the areas of wireless ommuniation and multifuntional
sen-sors with ommuniation and proessing apability have stimulated the development
and use of wirelesssensornetworks (WSNs) inmany dierentdomainssuhasthe
en-vironmental,medial,industrial,militaryelds andmanyotherwherehumanpresene
is not possible or desired [Boukerhe et al.,2007℄. A sensor node typially presents a
limitedsensingapability,but theoverallsensing apabilityanbeinreasedwhenthe
nodesare ombinedwithmanyother nodesformingaWSN. Forexample, ifagas leak
ours ina roomfull of gas ylinders and there isonly one sensor inthis room, itwill
onlybepossibletosaythatthereisaleakornot. Ontheotherhand,if aWSNisused,
with appropriate protools, it is possible not only to detet the leak, but to indiate
wherethe leakstartedandhowitevolved. A monitoringinthis way ansavelivesand
assets, and redueost insurane.
Sensor nodes are energy-onstrained devies and the energy onsumption is
generally assoiated with the amount of gathered data, sine ommuniation is
often the most expensive ativity in terms of energy. For that reason,
algo-rithms and protools designed for WSNs should onsider the energy onsumption
in their design [Olariuet al.,2004, AbdelSalamand Olariu,2009, Villasetal., 2010a,
Villas etal.,2011℄. Moreover, WSNs are data-driven networks that usually produe
a large amount of information that needs to be routed, often in a multihop fashion,
toward a sink node, whih works as a gateway to a monitoring enter. Given this
senario, routingplays animportantrole in the data gathering proess.
For more eient and eetive data gathering with a minimum use of
by making loal deisions [Chatzigiannakis etal., 2005, Chatzigiannakis etal., 2006,
Efthymiouetal., 2006, Villas etal.,2010b℄. For this, data aggregation and
spatio-temporaldata orrelation are eetive tehniques for savingenergy in WSNs. Due to
theinherentredundanyinrawdatagathered bysensornodes,in-networking
aggrega-tionandspatio-temporaldataorrelationanoftenbeusedtodereasethe
ommunia-tionostby eliminatingdataredundanyand forwardingonlyaggregatedinformation.
Sineminimalommuniationleads diretlyto energysavings, whih extendsthe
net-work lifetime, in-network data aggregation is a key tehnology to be supported by
WSNs.
Data aggregation has been used in WSNs with two purposes: (i)to take
advan-tage of the redundany and improve data auray [Shmid and Shossmaier, 2001,
Chakrabarty et al.,2002℄, and (ii) to redue the overall data tra and save
en-ergy [Krishnamahari etal., 2002℄. Nevertheless, urrent proposals have a high ost
toreate routing strutures aware of data aggregation and many of them do not
on-sider the spatio-temporal data orrelation. In addition, most proposals do not deal
with node failures and interruptions during a ommuniation, whih ause data loss
and donot guarantee delivery of the senseddata.
One of the main hallenges in routingalgorithms for WSNs ishow to guarantee
the delivery of the sensed data even inthe presene of node failures and interruptions
during ommuniation. These failures beome even more ritial when data
aggrega-tion is performed along the routing paths sine pakets with aggregated data ontain
informationfromvarioussouresand, wheneverone of thesepakets is losta
onsider-able amount of informationwill also be lost. For this reason, data aggregation aware
routingprotools shouldpresent a reliabledata transmission, through a fault-tolerant
routingmehanism.
1.2 Objetives
Themaingoalsofthis workare twofold. First,weprovideageneraldisussion fordata
aggregationandspatio-temporaldata orrelationproblems inWSNs. The seondgoal
istopropose, design, andevaluatethe performaneof dierent types ofalgorithmsfor
the problems of data aggregation and spatio-temporal data orrelation for WSNs. To
ahievethese goals, some seondary objetivesshould be aomplished:
and spatio-temporal data orrelation in WSNs, the following goals need to be
ahieved:
1.1. survey the state-of-the-art about the use of data aggregation and
spatio-temporal data orrelationin WSNs;
1.2. assess the arhitetures, models, and methods of data aggregation and
spatio-temporaldata orrelation identied inthe survey;
1.3. identify drawbaks ofurrent proposalsto propose new solutionsthat
over-ome the drawbaks of urrent proposals; and
2. For the seondmain goal,to propose dierentsolutionsfor the problems of data
aggregation and spatio-temporal data orrelation that are suitable for dierent
senarios ina WSN,the following goals need tobeahieved:
2.1. speifyanddesignalgorithmsthat onsiderthedatastruture,data
orrela-tions (spatial and temporal), the network topology and appliation
restri-tions;
2.2. propose a solution for the data aggregation problemto be used in medium
sale WSNs;
2.3. proposeasolutionforthe dataaggregationproblemtobeusedinlargesale
WSNs; and
2.4. analyze the performane of the proposed solutions.
1.3 Main Contributions
The main ontributions of this thesis are the designand development of fourdierent
solutions for data aggregation and spatio-temporaldata orrelation for WSNs, whih
we refer to as the DAARP, DDAARP, DST, and EAST algorithms, respetively. In
summary, wehave:
Data Aggregation Aware Routing Protool for WSNs (DAARP) is a novel
rea-tive data aggregation aware routing protool for WSNs. The main motivation
to design a new data aggregation aware routing protool is that the proposed
solutions in the literature present high ost to reate routing strutures aware
of data aggregation. The DAARP algorithmbuilds arouting struture with the
maximiz-along the ommuniation route, in a more reliableway, through a routing
fault-tolerant mehanism. Simulation results (presented in Setion 3.6) reveal that
DAARP has some keys aspets requiredby dataaggregation inWSNs suhas a
redued number messages for settingup a routingstruture, maximized number
of overlapping routes, high aggregation rate, and reliable data aggregation and
transmission. This algorithmis fully explainedin Chapter 3.
Dynami Data-Aggregation Aware Routing Protool for WSNs (DDAARP) is a
noveldynamidata-aggregationaware routingprotoolforWSNs,whihusesthe
sink node for proessing and onguration of routes aware of data aggregation.
The main motivation to design a dynami approah to reate dynami routing
strutures aware of data aggregation is that we have identied that the quality
of routing strutures reated by DAARP and the most algorithms in the
liter-ature depend on the order of events ourrene and one reated, these routes
are heldxed duringthe ourreneof events. Themainontributionisthatthe
routesreated byDDAARPdonotdependonthe orderofeventsourreneand
are not held xed during the ourrene of events suh as the DAARP.
Simula-tion results (presented in Setion4.6) revealthat DDAARP presents lowost in
terms of paket ontrol, improves the quality of the routing struture and
max-imizes data aggregation along the ommuniation route in a more reliable way,
through a routing fault-tolerane mehanism. This algorithm is fully explained
in Chapter 4.
Dynami and Salable Tree for WSNs (DST) is an eient data aggregation
solution that allows salable and dynami routing in WSNs, whih builds
rout-ing strutures with the shortest routes (inEulidean distane) that onnets all
sourenodes tothe sinknodemaximizingdataaggregationandreduing the
dis-tanetoonneteahsourenodetothesink. Also,theroutingstruturereated
does not depend on the event order. The main motivation to design the DST
was the lak of a solution in the literature salable, dynami and presents low
ost to reate routing strutures aware of data aggregation. DDAARP presents
goodresults, butitsuersfromsalabilityproblemsand beomesimpratialfor
large-salenetworks. Inaddition,thesinknodeneedstohaveaglobalknowledge
of the network. Simulations results (presented in Setion 5.6) reveal that DST
presentsalowostinterms ofpaketontrol, maximizesaggregationpointsand
Eient Data Colletion Aware of Spatio-Temporal Correlation for WSNs
(EAST)isanalgorithmforenergy-awaredataforwardinginWSNsthattakesfull
advantage of both spatial and temporal orrelation mehanisms to save energy
while still maintaining real-time, aurate data report towards the sink node.
The main motivation to design the EAST is that most of the urrent spatial
and/or temporal orrelation algorithms do not onsider the energy dissipation
during data olletionto better hoose the representative nodes. Also, these
so-lutions present a high number of ontrolmessages and donot exploit eiently
the spatio-temporal orrelation nor their dynamiity. The main ontribution
is an energy-aware spatio-temporal orrelation mehanism in whih nodes that
deteted the same event are dynamially grouped in orrelated regions and a
representative node is seleted at eah orrelation region for observing the
phe-nomenon. The entire region of sensors per event is eetively a set of
repre-sentative nodes performingthe task of data olletionand temporalorrelation.
Simulationresults(presented inSetion6.4)learlyshowthatbyusingboth
spa-tialandtemporalorrelation,theinformationabouttheeventanbesensedwith
a highauraywhile stillsavingthe residualenergy ofnodes. This algorithmis
fully explained inChapter 6.
1.4 Organization of the Thesis
This thesis is divided into seven hapters. Chapter 2 provides an overview about
wirelesssensornetworks,in-networkaggregationand spatio-temporaldata orrelation.
Thehapterintrodueswirelesssensornetworks,disusses in-networkdataaggregation
andpresentsthemaintehniquestoperformdataaggregation. Inaddition,itdisusses
spatio-temporaldata orrelation and provides an overview of the existing approahes
that exploit spatio-temporaldata orrelation.
In the seond part of this work, omposed of Chapters 3, 4,5 and 6,wepropose
and explain the DAARP, DDAARP, DSTand EAST algorithms,respetively. In eah
hapter,the performane of the proposed solutionis evaluatedthrough simulations.
Finally,inChapter 7,wepresentsome nalremarksabout thestudiedproblems,
Bakground
Thishapterpresentsthetheoretialfoundationforthiswork. Thishapterisorganized
as follows: Setion 2.1 introdues wireless sensor networks, Setion 2.2 disusses
in-network data aggregation and presents the related work, and Setion 2.3 disusses
spatio-temporaldata orrelation and provides an overview of the existing approahes
whih exploit spatio-temporaldata orrelation.
2.1 Wireless Sensor Networks
Wireless Sensor Networks (WSNs) [Akyildizet al.,2002, Romer and Mattern,2004,
Boukerhe etal.,2007, Anastasiet al.,2009℄ an be dened as a ooperative network
of small, battery-operated, wireless sensor nodes whose main goal is twofold: to
monitor their surroundings for loal data and to forward the gathered data toward
a sink node using typially multihop ommuniation. This sink node will then be
responsible for proessing all of the reeived data from several soure nodes and
reporting them to a monitoring faility (Figure 2.1). This type of network has
beome popular due to its appliability that inludes several areas suh as
envi-ronment, homeland seurity, industry, domestis, agriulture, meteorology, health,
spae, military and many other appliations that an be ritial to save lives and
assets [Younis etal., 2006, Anastasiet al.,2009, Villaset al.,2010b℄. Several physial
propertiesanbemonitored,inludingtemperature, humidity,pressure, ambientlight,
sound, vibration, and motion.
One of the main limitationsof the WSNsis the battery-operated natureof their
sensor nodes, whih makes this kind of network highly energy-onstrained. A
sim-ple solution to this problem ould be the periodi replaement of the node battery.
net-Figure 2.1. Data routing inWSNs [Oliveira etal., 2009℄.
work or beause the sensor nodes may be inaessible in some appliations suh as
monitoring volanoes or spae exploration. For that reason, algorithms and
pro-tools designed for WSNs should onsider the energy onsumption in their
onep-tion [Olariu etal., 2004, AbdelSalamand Olariu, 2009, Villaset al.,2011℄.
For more eient data gathering with a minimum use of limited
re-soures, sensors should be ongured to report data more intelligently by
making loal deisions [Chatzigiannakiset al.,2005, Chatzigiannakis etal., 2006,
Efthymiouetal., 2006,Villaset al.,2010b℄. Dataaggregation 1
andspatio-temporal 2
,
3data orrelation are possible tehniques for loal deision-making, whih will be
pre-sented inthe following setions. Suhstrategies help tomaximize energyonservation
inanappliation-speisensor network.
These tehniques have been exploited in the literature suh as data
aggrega-tion [Krishnamahari et al.,2002, Chandrakasan et al.,2002, Nakamura etal., 2006,
Fanetal., 2006, Nakamura et al.,2009℄, spatial orrelation [Akyildiz etal., 2004,
Yoonand Shahabi, 2005, Liuet al.,2007b, Leet al.,2008, Yuan and Chen, 2009℄and
temporalorrelation [Minand Chung, 2010, Pham et al.,2010℄. Nevertheless, urrent
proposals have a high ost to reate routingstrutures aware of data aggregationand
manyof themdoesnot dealnodesfailures andinterruptionsinommuniations,whih
auses loss of data and does not guarantee delivery of the sensed data. In addition,
these solutions not only introdue delays in data transmissions but also lead to the
reeption ofoutdated informationby the sink node.
1
Data Aggregationeliminates inherentredundany in rawdata gatheredbythe sensornodes
andforwardingonlysmalleraggregatedinformation.
2
Spatial orrelation: the hange pattern of the data sensed by nearby nodes is expeted to
be the same or similar. Thus, exploit the spatial data orrelation an eliminate the similars data
reporting.
3
Temporal orrelation: the hange patternin readings ofasensor nodeand gathereddatais
usuallysimilar overashort-timeperiod. Due tothe natureof thephysial phenomenons, thereis a
2.2 In-network Data Aggregation
In the ontext of WSNs, in-network data aggregation refers to dierent tehniques
to forward data pakets toward the sink node. During this proess, nodes ombine
data olleted by dierent soures, i.e., by fusing sensor readings related to the same
event or physial quantity, or by loally proessing raw data before its transmission.
A key omponent of in-network data aggregation is the design of a data aggregation
aware routing protool, whih determines where the aggregation shall be performed.
Data aggregation requires a forwarding paradigm that is dierent from the lassi
routing,whihtypiallyinvolvestheshortest path inrelationtosome speimetri
to forward data toward the sink node. Dierently from the lassi approah in data
aggregation aware routing protools, hooses the node as next hop based on their
proximity in the topology that maximizes the overlap of routes in order to promote
in-network data aggregation.
Before lassifying the literature on solutions aware data aggregation, rst we
illustrate the importaneof oupling between routingand data aggregation in WSNs.
As depited in Figure 2.2, assume that the routing struture of data olletion in a
wsn isainvertedmultiasttreerootedatthesink(nodeA).Theonept ofin-network
data aggregation an be illustrated as follow.
Figure 2.2. DataAggregation AwareRouting
Leteahof thenodes in
{
E, G
andH
}
inthesensingeld1and{
F, I, J
andK
}
in the sensing eld2 generate one raw sensory paket of size
r
ands
bits respetively.The arrows in the gure indiate the data gathering struture. For example, data
olleted by the sensing nodes
{
E, G
andH
}
in the sensing eld 1 will be sent tothe sink via nodes
C
andB
. As olleted data from physially proximate nodes areThe resultis that the size of the outgoingpaket from
E
,R
bits, willbe less than thesummation of all the inomingpakets (inluding its own) and often larger than any
ofthe individualinomingpakets. The amount of the redutiondepends onthe data
orrelation as speied by the appliation. For example, in the extreme ase where
there is no data orrelation, we have
R
=3
r
(inluding the data sensed by nodeE
)asno data redution an beahieved. On the other hand, if the desired result is, say,
simplyaverage of the measurements,
R
=r
. However, most appliations whih datahasaertaindegreeoforrelationwillsatisfy
k < K <
3
k
. ThenodeB
aggregatestheolleted data in the sensing eld 1 and 2. The result is that the size of the outgoing
paketfrom
B
,T
bits,willbelessthanthesummationoftheinomingpaketsofR
andS
bits and often larger than the lower individual inoming pakets of sizemin
(
R, S
)
bits, ie,
min
(
R, S
)
< T < R
+
S
.Based in the above example, a key fator in the proess of data aggregation is
the routing sheme, whih determines where the aggregation shall be performed. For
in-network data aggregation to be realized, some form of synhronization is needed
among the nodes. Typially, in these algorithms, a node usually does not send data
as soon as it is available sine waiting for data from neighboring nodes may lead to
better data aggregation opportunities. This in turn, will improve the performane of
thealgorithmandsaveenergy. Threemaintimingstrategiesare foundintheliterature
[Solis and Obrazka, 2004, Hu etal.,2005℄:
Periodi simple aggregation: requires eah node towait for a pre-dened period
of time while aggregating all reeived data paket and, then, forwards a single
paket with the result of the aggregation.
Periodi per-hop aggregation: quite similar to the previous approah, but the
aggregated data paket is transmitted as soon as the node hears from all of its
hildren. This approah requires eah node to know the number of its hildren.
In addition,a timeout may be used for the ase of some hildren's paket being
lost.
Periodi per-hop adjusted aggregation: adjusts the transmission time of a node
aording to this node's positionin the gathering tree.
Note that the hoie of the timing strategy strongly aets the aggregation rate
in-networkandlatenytoreportdataolleted. Ifthewaitingtimeinaggregationpoint
2.2.1 Routing Sheme Aware Data Aggregation
In-networkdataaggregationplaysanimportantroleinenergyonstrainedWSNssine
dataorrelationisexploitedandaggregationisperformedatintermediatenodes
redu-ing size and the number of messages exhanged aross the network. Indata gathering
basedappliations,aonsiderablenumberofommuniationpakets anbereduedby
in-networkaggregation, resultinginalongernetwork lifetime. The problemof
obtain-ing the optimalaggregation tree is known to be
N P −
Hard
[Al-Karaki etal., 2004℄,whih isequivalent tothe Steiner tree problem[Krishnamahari etal.,2002℄.
Denition 1 (Steiner Tree) given a network represented by a graph
G
= (
V, E
)
,where
V
=
{
v
1
, v
2
, . . . , v
n}
istheset of sensornodes,E
istheset of edges representingtheonnetionsamongthenodes,i.e.,
h
i, j
i ∈
E
iv
i
reahesv
j
,andw
(
e
)
istheostofedge
e
,aminimalosttreeistobebuiltthatspansallsourenodesS
=
{
s
1
, s
2
, . . . , s
m}
,S
⊆
V
, and the sinknodes
0
. The ost of the resulting Steiner tree (W
) is the sum ofthe osts of its edges. This problem is a well-known NP-hard problem.
In the literature, there are dierent heuristis to the Steiner tree
prob-lem, some of them present
1
.
598
approximation fator [Hougardy and Prömel,1999,Robins and Zelikovsky, 2000℄. However, thosesolutionsare not aordableto
resoure-onstrained networks, suh as WSNs, sine their distributed implementation requires
alarge amountofmessages tosetuptheroutingtree,whihonsequentlyleads tohigh
energy onsumption.
Some researh eorts have also been made to develop routing algorithms for
WSNs. Table2.1presentsasummaryofthe basiharateristisofthemain proposed
routing protools for WSNs. In this work, we lassify these proposals into three
at-egories: tree-based, luster-based, and struture-less algorithms. In the next setions,
wewillbriey review these protools and their strutures.
2.2.1.1 Tree-Based Approahes
Protools in this family [Al-Karakiand Kamal,2004, Akkaya and Younis,2005,
Fasoloet al.,2007℄ are usually based on a hierarhial organization of the nodes in
the network. In fat, the simplest way to aggregate data owing from the soures to
the sinknode istoeletsomespeial nodes thatworkasaggregationpointsand dene
a preferred diretiontobe followed when forwarding data.
Intheseprotools,atreestrutureisonstrutedrstandthenusedlatertoeither
Table 2.1. Summaryofthebasiharateristisofsomedataaggregationaware routing protools Sheme Route Stru-ture Objetive Aggregation Nodes Overhead Salability Drawbak LEACH Cluster-based Maximize life-time
Clusterheads Medium Low
Low
salabil-ity
SPT Tree Shortest-Path
Opportunisti Low High Data redun-dany GIT
Treewithpath
sharing
Minimizetotal
energyost
Intermediate
nodes
VeryHigh VeryLow Highost
CNS Tree
Aggregateloser
tothesink
Aggregatornode High
Medium Only one aggregation point InFRA Tree-based lus-ter Maximize over-laproutes Clusterheads andintermediate nodes
VeryHigh
Low
Low
salabil-ity and high
ost DAARP Tree-based lus-ter Maximize over-laproutes Clusterheads andintermediate nodes
Medium Medium Statiroutes
DDAARP Tree-based lus-ter Maximize over-laproutes Clusterheads andintermediate nodes Low Medium Requires global knowl-edge DST Based on straight line segmentsand luster Maximize
over-lap routes and
minimize
over-head
Clusterheads
andintermediate
nodes
VeryLow VeryHigh
Requires posi-tion informa-tion DAA Anyast Minimize over-head Intermediate nodes
VeryLow VeryHigh
Notallpakets
maybe
aggre-gated
ofthe tree. This nodethenaggregatesallreeived datawithitsowndataandforwards
only one paket to its neighbor that is lower in the tree. However, this approah has
somedrawbaks. Forinstane, whena paketis lostat aertainlevelof the tree (e.g.,
due tohannel impairments), data from the whole sub tree willbelost aswell. Thus,
tree-based approahes require a mehanism for fault tolerane to reliably forward the
aggregateddata.
Despite the potentially high ost of maintaining a hierarhial struture in
dy-nami networks and the sare robustness of the system in ase of link/devie
fail-ures, these approahes are still partiularlysuitablefor designing optimal aggregation
funtions and performing eient energy management. For instane, there are some
proposed solutions [III et al.,2007, Villaset al.,2010a℄ where the sink node organizes
routing paths to evenly and optimally distribute the energy onsumption while still
favoring the aggregation of data atthe intermediate nodes.
Inmost ases,tree-based protools buildatraditionalshortest path routingtree.
For instane, the Shortest Path Tree (SPT) algorithm [Krishnamahari etal., 2002℄
uses a very simple strategy to build a routing tree in a distributed fashion. In this
approah,everynode thatdetets aneventreportsitsolleted informationby usinga
shortest path tothe sink node. Data aggregation ours whenever paths overlap
In these ases, data an be opportunistially aggregated when they meet at any
intermediate node. Based on Direted Diusion, the Greedy Inremental Tree
(GIT)[Intanagonwiwat etal.,2002℄approahwasproposed. TheGITalgorithm
estab-lishes anenergy-eient pathand greedily attahes other souresontothe established
path. In the GIT strategy, when the rst event is deteted, nodes send their
infor-mation as in the SPT algorithm and, for every new event, the information is routed
using the shortest path to the urrent tree. There is a new aggregation point every
time a new branh is reated. Some pratial issues make GIT not appropriate in
WSNs [Nakamura et al.,2006℄. For example, eah node needs to know the shortest
path toallnodes inthenetwork. Theommuniationosttoreatethisinfrastruture
is
O
(
n
2
)
,where
n
isthe number ofnodes. Furthermore, thespae needed tostore thisinformation at eah node is
O
(
D n
)
, whereD
is the number of hops in the shortestpath onneting the farthest node
v
∈
V
tothe sink node (network diameter). Afterthe initialphasethe algorithmneeds
O
(
m n
)
messages tobuildtheroutingtree, wherem
isthe numberof soure nodes.Another interesting solution is the Center at Nearest Soure (CNS)
algo-rithm [Krishnamahari et al.,2002℄. In CNS, every node that detets an event sends
itsinformationtoaspei node, alledthe aggregator,by usinga shortest path. The
aggregatoristhe losestnode tothe sink(inhops)that detets anevent. CNS redues
the amount of data sent to the sink in relation to the lassial approahes, but the
overhead inCNS is highlydependent onthe events' ourrene positions. In senarios
with many events ourring simultaneously,CNS has ahigh ost to hange the
aggre-gatornode. Inthisalgorithm,data redundanyisonlyreduedwhenitisalreadylose
to the sink node.
2.2.1.2 Cluster-Based Approahes
Similarly to tree-based approahes, luster-based shemes [Chandrakasan etal.,2002,
Nakamura etal., 2006, Villasetal., 2009, Villaset al.,2010a℄ alsoonsist of a
hierar-hial organization of the network. However, in this approah, nodes are subdivided
into lusters. Moreover, speial nodes, referred to as luster-heads, are eleted to
ag-gregate data loallyand transmitthe result of suhan aggregationto the sink node.
In the Low-Energy Adaptive Clustering Hierarhy (LEACH)
algo-rithm [Chandrakasan et al.,2002℄, lustered strutures are exploited to perform
data aggregation. In this algorithm, luster-heads an at as aggregation points and
LEACH-based algorithms assume that the sink an be reahed by any node in only
onehop, whihlimitsthe size of the networkfor whihsuhprotools an be used. In
addition,in senarios where the data an not be perfetly aggregated, LEACH-based
protools donot neessarily have signiantadvantage sine the luster-heads have to
send many pakets to the sink using ahigh transmission power.
The Information Fusion-based Role Assignment (InFRA)
algo-rithm [Nakamura et al.,2006℄ builds a luster for eah event inluding only those
nodes that were able to detet it. Then, luster-heads merge the data within the
luster and send the result to the sink node. The InFRA algorithm aims to build
the shortest path tree that maximizes the information fusion. One lusters are
formed, luster-heads hoose the shortest path (to the sink node) that maximizes
the informationfusion with already formed paths/lusters [Nakamura et al.,2006℄. A
disadvantage of the InFRA algorithm is that for eah new event that arises in the
network, the informationabout the event must be ooded throughout the network to
inform other nodes about its ourrene and to update the paths from the already
existing luster-headstothe sinknode. This proedurelimitsInFRA's salability.
Another interesting solution is the Data Aggregation Aware Routing Protool
(DAARP) [Villaset al.,2009℄. For eah event this algorithm performs the lustering
ofnodes that deteted the sameevent,as wellas the eletionof aluster-head. Then,
luster-heads merge data within the luster and send the result to the sink node.
Afterthe luster-head formation,routes are reated by seletingnodes in the shortest
path (in hops) to the nearest node that is part of an existing routing infrastruture
in whih this node will be an aggregation point. The DAARP routing infrastruture
tends to maximize the aggregation points and uses fewer ontrolpakets to build the
paths. Dierent fromInFRA,DAARP doesnot ood amessagetothe whole network
whenevera new event ours. DAARP is not feasiblefor senarios with long duration
events beause the routes are stati, whih quikly onsumes the energy of the nodes
that are part of the routingstruture.
The Dynami Data Aggregation Aware Routing Protool
(DDAARP)[Villasetal., 2010a℄adds animprovementoverDAARP.IntheDDAARP
algorithm,the routes are omputed at the sink node and donot depend on the order
of events. Routes reated by DDAARP are not kept xed throughout the duration
of events, i.e., routes may hange when neessary. The drawbak of this proposal
is that pakets ontaining information from nodes tend to inrease their size at the
information olleting stage and this solution beomes impratial for large-sale
2.2.1.3 Struture-Less Approahes
Few algorithmsfor routing aware of data aggregation have been proposed that use a
struture-lessapproah. The Data-Aware Anyast (DAA)algorithm[Fanetal., 2006℄,
a struture-less data aggregation algorithm, uses anyast to forward pakets to
one-hop neighbors that have pakets to be aggregated. It involvesmehanismsto inrease
the hane of pakets meeting at the same node (spatial aggregation) at the same
time (temporalaggregation). Sinetheapproahdoesnot guarantee aggregationofall
pakets, the ost of transmitting pakets with no aggregation inreases with the size
of the network. In addition,pakets that are unable to be aggregated willnot benet
from the energy savings ahieved by eliminatingthe ontroloverhead.
The Dynami and Salable Tree (DST) algorithm [Villas etal.,2010b℄ aims to
build a routing tree with the shortest routes (in Eulidean distane) that onnets
all soure nodes to the sink node, maximizing data aggregation while reduing the
distane onneting eah oordinator node to the sink. Routes are based on straight
line segments, whih are omputed by the oordinator nodes. The reated paths do
not depend on the event order. Similar to all approahes that do not exploit the
spatial orrelation, DST does not show a good performane in senarios where many
nodes detet the same event, sine nodes that reportinformationabout the event an
onsume their energy quikly.
2.3 Exploiting Spatio-Temporal Correlation
In this setion, we present the benets of exploiting spatio-temporal data orrelation
in WSNs. We also disuss some of the existing approahes and algorithmsthat take
advantage of spatio-temporalorrelation in WSNs. Table 2.2 presents a summary of
thebasiharateristisofthemainproposedspatialand/ortemporaldataorrelations
algorithmsfor WSNs.
In the urrent literature, we an nd three main ategories of data orrelation
protools: (i) spatial orrelation; (ii) temporal orrelation and (iii) spatio-temporal
orrelation. Inthe following,wepresent someof theseprotoolsaswellas thebenets
of exploiting spatial/temporaldata orrelation inWSNs.
2.3.1 Spatial Correlation
Table 2.2. Summary of the basi harateristis of the main proposed spatial
and/ortemporaldataorrelations algorithms for WSNs
Sheme Route Struture Objetive Spatial Correl. Temporal Correl. Overhead Salability Drawbak EEDC Singlehop Eliminate on-troloverhead Yes No
VeryLow VeryLow
Centralized andsingle-hop network CAG Tree-based luster Eliminatedata redundany Yes No
VeryHigh
Medium Maintenane data-entri GSC Tree-based luster Eliminatedata redundany Yes No High Low
Isnotapplied
to multi-hop
members
SBR Tree-based
Eliminatedata
redundany
No Yes Medium
High Sink node an reeive outdated infor-mation SCCS Tree-based luster Eliminatedata redundany
Yes Yes Medium
High Sink node an reeive outdated infor-mation EAST Based on straightline segments andluster Maximize
over-lap routes and
minimize
on-troloverhead
Yes Yes
VeryLow VeryHigh
Requires
position
infor-mation
Yoonand Shahabi, 2005, Yuan and Chen, 2009, Nakamura etal., 2009,
Villasetal., 2009, Villas etal.,2010a, Villaset al.,2010b, Villaset al.,2011℄. The
eletednodeisthen responsible forreeivingalltheeventnotiationsand forwarding
them toward the sink node. The energy onsumption of the nodes that detet events
isgreaterthanthe othernetworknodes (seeFigure2.3(a)). Thisoursbeausenodes
within the group (nodes that detet events) onsume a great deal of energy reeiving
and forwardingdata pakets from their neighbors, besides their own notiations.
As aninitialmotivation,Figure2.3presents the energy onsumption inthe
pro-ess of data olletion in a WSN of two dierent routing approahes when the sink
node,loatedatposition(0,0),reeivesdatafromadeteted eventthathas aradiusof
70mandisloatedatposition(600,600). Therstapproah(Figure2.3(a))isasimple
methodfordataolletionwhereallnodesthatdetetedtheeventsendthesenseddata
toward the sink node. The seond approah (Figure 2.3(b)) is a more sophistiated
strategythatusesspatialorrelationtosaveenergy. Inthisase,onlyasubsetofnodes
that deteted the event sends sensory data to the sink node. In both senarios, the
notiationofthedetetedeventwasperformedateahseond,andtheeventduration
wasofonly10m. Therstapproah(Figure2.3(a))sends32157notiations,whereas
the seondapproah (Figure2.3(b)) sends only5667 notiations.
The dierenebetween the two approahes is notableand, by using spatial
or-relation,the seondapproahwasable tosavealarge amountofenergy,extendingthe
overall network lifetime.
0
100
200
300
400
500
600
700
Meters (m)
0
100
200
300
400
500
600
700
Meters (m)
0
20
40
60
80
100
Consumption Energy (J)
(a)
0
100
200
300
400
500
600
700
Meters (m)
0
100
200
300
400
500
600
700
Meters (m)
0
20
40
60
80
100
Consumption Energy (J)
(b)
Figure2.3. Energy onsumptionofnodesduringthedataolletionwhenusing
(a)alassialapproah;andwhenusing(b)aspatialorrelationbasedapproah.
mation. In this ase, instead of having all sensor nodes reporting the same data, it is
more eient to hoose a few representative nodes to notify the sink node about the
detetedevent(seeFigure2.3(b)). Arepresentativenodereportstheeventinformation
of a given area on behalf of a group of nodes that ollets similar information in the
same area.
Akyildiz et al. [Akyildiz etal., 2004℄ studied the relation between reliability of
event detetion and spatial loation of the sensor nodes in the event area. Their
solu-tion estimates the numberof sensor nodes (representative nodes) required tosend the
deteted event to the sink in order to have reliable event information. Eah
represen-tative node represents a spatially orrelated group of nodes. Although their solution
ahieves overall energy gain, it fails to onsider the remaining energy during the
se-letion of the representative nodes an assumption that should not be negleted in a
WSNbeauseofhardwareonstraints. Thus, ifarepresentativenodeworksinthe
or-relation regionfor a long periodof time, itwill spend more energy due tothe number
of transmitted messages omparedto the other nodes.
YoonandShahabi[Yoonand Shahabi, 2005℄proposedanew mehanismfor
spa-tial orrelation in WSNs. The proposed mehanism, alled Clustered Aggregation
Tehnique (CAG), reates lusters of nodes with similar sensing values and only a
node inside the luster noties its reading to the Sink node whereas the other nodes
ignore their readings. The CAG algorithm is divided into two phases: query and
response. Inthe queryphase,the data-entrilustersare reated aordingtoa
belong to the same luster. In the seond phase (response phase), just one node per
luster (luster-head) sends its sensed value to the sink node notifying the deteted
event. The authorsshowed that the proposed mehanism an reduesigniantly the
number of transmitted messages during the data olletion. However, during the rst
phase, the CAG algorithmuses a ooding-based protooltodisseminate the query to
allsensornodes, whihis not needed inmost senarios. Moreover, the maintenaneof
the data-entri lustersremains adiult problem[Boukerhe etal., 2003℄.
Liu et al. [Liu etal., 2007a℄ proposed another lustering algorithm, named
Energy-Eient Data Colletion framework (EEDC), to exploit spatial data
orre-lation. They onsider that nodes olletdata ontinuously and are one-hop onneted
to the sink node or to a enter node. The algorithm was designed to be exeuted at
thesink node, sinethis node has the entire datanetwork information. Thealgorithm
reateslustersof nodesthatare spatiallyorrelated. Also,the sinknode managesthe
lusterformation dynamiallyinorder to reetenvironmental hanges. The primary
limitation of that sheme is the assumption of the single-hop ommuniation. This
assumption is impratial in a distributed system and diult to have in large-sale
wirelesssensor networks. Anotherdisadvantage isthelustering algorithmthatis
en-tralizedatthesinknode. Beause ofthis, allnetworkdataneeds tobesent tothe sink
node, whih willstore and proess a great amount of data.
Shah etal.[Shahand Bozyigit,2007℄proposed anew mehanism forspatial
or-relation in WSNs, named Gridiron Spatial Correlation (GSC). The GSC is adaptive
toahievetherequired reliabilityby dynamiallyhangingthe orrelationregion. The
orrelation regions are formed as squared retangles and nodes lying in the retangle
areassumed tobespatiallyorrelated. Cluster-headidentiestheredundantandlose
soures in its viinity and turns o the ativity of nodes by onsidering their energy
levelandloseness asriterion. ThelimitationofGSC isthe ontrolmehanismwhih
is not applied to multi-hop members, ie, it only works well for senarios where the
event radius issmaller than the ommuniationradius of the luster-head.
2.3.2 Temporal Correlation
Sensorreadings about the environment are typially periodi; onsequently, the
time-orderedsequene ofsenseddataonstitutesatimeseries(seeFigure2.4(a)). Duetothe
nature of the physialphenomenon, there is a signiant temporal orrelation among
eah onseutive observation of a sensor node and gathered data is usually similar
